release aidev BREAKING SIG 3/5

Aider v0.83.0: drops Python 3.9, Gemini 2.5 Pro aliases move to 05-06 preview

Aider dropped support for Python 3.9 — an environment-level break for anyone still on it. It also repointed the gemini and gemini-2.5-pro aliases at the 05-06 preview models, added qwen3-235b support, and enabled thinking_tokens / reasoning_effort for OpenRouter models.

PUBLISHED2025-05-09
OBSERVED2026-08-11
AGE1y
SOURCES1
  • Python 3.9 support dropped.
  • Added gemini-2.5-pro-preview-05-06 and qwen3-235b models.
  • Gemini aliases (gemini, gemini-2.5-pro) repointed to the 05-06 previews; those previews marked overeager by default.
  • Default weak model for Gemini 2.5 Pro is now gemini/gemini-2.5-flash-preview-04-17; reasoning_effort enabled for Gemini 2.5 Flash.
  • Corrected gemini-2.5-pro-exp-03-25 settings — it does not support thinking_budget.
  • thinking_tokens and reasoning_effort now supported for OpenRouter models; model params (context window, pricing) fetched from OpenRouter directly; cost calculation uses litellm.completion_cost where available.
  • New --attribute-co-authored-by adds a co-author trailer; explicit --attribute-author / --attribute-committer override it.
  • New --disable-playwright flag; aider scrape and the GUI use Playwright when available and honour the flag.
  • New --shell-completions argument to emit bash/zsh completion scripts.
  • Repo-map support for OCaml and OCaml interface files; repo map shows token progress and the file being processed.
  • Windows linter execution fixed via oslex argument quoting; same for cross-platform shell command display.
  • /ask instructs the model to elide unchanged code.
  • Fix: read-only file status not preserved by some commands (e.g. /drop).
  • aider-args prints a sample YAML config when given no arguments.
  • Total tokens sent/received now tracked in benchmark stats.
  • Spinner added while waiting for the stream to start, extended to non-streaming responses.

COMMUNITY

No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.